Wavelet Fusion Network for Signal Denoising

Inertial sensors are fundamental motion-sensing components widely utilized in navigation, aerospace, and consumer electronics. However, their inherent noise significantly compromises signal quality and the reliability of downstream applications. Traditional wavelet denoising methods rely on manually selecting fixed wavelet basis functions, struggling to adapt to complex and varying noise types and signal characteristics. This paper proposes an inertial signal enhancement method based on weighted wavelet fusion, termed the Wavelet Fusion Network (WFNet). WFNet dynamically generates adaptive synthesized wavelets via a deep learning model, overcoming the limitations of single wavelet basis functions. To enhance the deep learning architecture's ability to perceive the distinct characteristics of different wavelet bases, we introduce a Rényi entropy based Class Representation Regularizer (CRR). This regularizer structures the weight matrix within the wavelet classifier, encouraging distinct and informative representations for each candidate wavelet, thereby sharpening the model's sensitivity to unique wavelet properties. Experiments demonstrate that WFNet significantly outperforms existing methods in both static evaluations like Allan variance analysis and dynamic tasks such as attitude estimation and trajectory reconstruction. This research presents a new paradigm for adaptive wavelet basis design and achieves deep synergy between model driven and data driven approaches through its weighted fusion strategy, offering an effective solution for signal enhancement in complex noise environments.